New Dataset Benchmarks Training Strategies for AI Models Predicting Blood Flow in Aneurysms

Researchers have created a comprehensive benchmark dataset to evaluate how machine learning models trained on synthetic variations of medical geometries can better predict physical fields across real patient cases. The work addresses a practical challenge in scientific computing: building faster surrogate models for complex simulations while working with limited real-world data.

The Dataset and Problem#

AneumoBench pairs 401 source aneurysm geometries with 9,693 locally edited variants, alongside complete computational fluid dynamics (CFD) simulations for each. The dataset encompasses 80,752 steady-state velocity and pressure cases computed under eight different inlet flow conditions, plus 9,715 time-varying sequences capturing velocity, pressure, and wall shear stress over cardiac cycles. Each sequence runs for a full heartbeat with 100 frames recorded at 0.01-second intervals.

The core question the benchmark addresses: does training on edited variants of limited geometries help models generalize to unseen patient-specific cases? And if so, what's the best way to allocate these synthetic variants across source geometries?

Training Approaches Compared#

Researchers evaluated nine different machine learning architectures using three distinct strategies on geometries withheld from training:

• Source-only training: Models trained exclusively on the original 401 aneurysm cases
• Descendant training: Models trained on locally edited variants
• Two-stage transfer: Pretraining on descendants followed by fine-tuning on source geometries

Results showed that two-stage training reduced prediction errors for steady flow fields and certain wall shear stress metrics compared to source-only approaches. The GraphSAGE architecture showed particular sensitivity to training strategy, benefiting from descendant data and—when given a fixed total number of variants—from spreading them across more diverse source geometries rather than concentrating them on fewer.

Important Nuances in Results#

The findings reveal complexities that matter for clinical applications. While descendant pretraining improved some error metrics over short time horizons, improvements to wall shear stress predictions did not consistently hold over longer forecasts (96 time steps). Additionally, lower prediction error over an entire sequence did not guarantee better performance on clinically relevant measures—such as identifying high-stress hotspots where aneurysms are most likely to rupture.